v0.2.1 — Long-running process visibility

Status: active package release on the v0.2 architecture line.

Version source: pyproject.toml (version = "0.2.1").

---

What changed

Non-architectural ergonomics added to full_llm_pipeline.ipynb. The tokenizer, model family, training objectives, dataset registry, and checkpoint formats stay on the v0.2.0 architecture line.

• Quiet pre-flight helpers stay visible. Dataset disk-size scans and load_dataset / load_from_disk calls are wrapped in Heartbeat(...), so blocked metadata or local Arrow loads emit keep-alive messages instead of going silent.
• Large list scans show progress. Text fingerprinting, train_texts character counting, and effective-mix counting use tqdm where a full-corpus run can spend noticeable wall time.
• Model setup and restore have heartbeat coverage. Model construction, parameter counting, checkpoint model rebuild, and checkpoint weight loading emit keep-alive messages if production configs run long.
• Speculative decoding reports progress. The demo adds a token progress bar plus heartbeat status for larger max_new_tokens runs.

Verification

• Notebook JSON validated with jq empty full_llm_pipeline.ipynb.
• Notebook code cells compiled with uv run python.
• uv run python -c "import download_data" passed.